基于深度学习的肺结节良恶性判别模型在靶扫描CT数据的效能验证  

The effectiveness of deep learning techniques in differentiating benign and malignant lung nodules on target CT scans

作　　者：陶广昱 叶剑定[1] 叶晓丹[1] 毛丽 虞凌明[1] 周振 李秀丽 Tao Guangyu;Ye Jianding;Ye Xiaodan;Mao Li;Yu Lingming;Zhou Zhen;Li Xiuli(Department of Radiology,Shanghai Chest Hospital,Shanghai Jiaotong University,Shanghai 200030,China;Deepwise Artificial Intelligence Lab,Beijing 100080,China)

机构地区：[1]上海市胸科医院上海交通大学附属胸科医院放射科,200030 [2]深睿医疗人工智能研究院,北京100080

出　　处：《中华放射学杂志》2019年第11期952-956,共5页Chinese Journal of Radiology

基　　金：上海市科学技术委员会"科技创新行动计划"高新项目(18511102902);上海交通大学"医工交叉研究基金"(YG2017QN66);上海市卫生和计划生育委员会科研课题(20184Y0219);上海市科学技术委员会西医引导类项目(19411965200);上海市数字媒体处理与传输重点实验室开放课题(STCSM18DZ2270700);上海市卫生健康委先进适宜技术推广项目(2019SY063)。

摘　　要：评估使用常规肺CT扫描数据训练的肺结节良恶性判别深度学习模型,在肺结节靶扫描CT图像上的良恶性分类效能.方法回顾性分析上海市胸科医院2016年1月至2018年12月间胸部CT扫描发现肺结节并行手术切除的患者923例,共搜集有病理检测报告且在常规扫描和靶扫描数据集上可对应的结节969个.使用基于常规扫描CT数据训练的深度学习良恶性分类模型,对于本研究中搜集的常规扫描和靶扫描数据集进行评测,评估指标包含两者的曲线下面积(AUC)、准确率、敏感度和特异度,并使用综合判别改善指数及Delong测试比较两者的性能差异.结果常规扫描数据集上的曲线下面积、准确率、敏感度和特异度分别为0.80、82.0%、86.0%和56.6%,靶扫描数据集上的曲线下面积、准确率、敏感度和特异度分别为0.84、85.0%、88.8%和60.5%.综合判别改善指数为0.056,差异有统计学意义(Z检验,P<0.05),且ROC曲线下面积差异有统计学意义(Delong检验,P=0.01).结论基于常规肺CT扫描数据训练的深度学习肺结节良恶性分类模型,在肺部靶扫描数据上可以取得更好的诊断效能.Objective To evaluate the effectiveness of deep learning model trained on routine CT scans when identity the malignant and benign lung nodule on target CT scans dataset.Methods This retrospective study enrolled 923 patients with lung nodules found by chest CT scan in Shanghai Chest Hospital from January 2016 to December 2018.A total of 969 nodules with pathological report were analyzed.The deep learning based pulmonary malignant prediction method in a fine-grained classification manner was used to make the prediction,and the AUC(the area under the curve),accuracy,sensitivity and specificity of routine CT scans and target CT scans were compared,and Delong test and IDI(Integrated Discrimination Improvement)were employed to provide statistical results.Furthermore,statistical methods were used to investigate the differences between the benign and malignant classification of nodules on routine CT and on target CT.Results In the benign and malignant discrimination task,AUC,accuracy,sensitivity and specificity on the routine scans were 0.81,82.0%,86.0% and 56.6% respectively,while the AUC,accuracy,sensitivity and specificity on the target scans were 0.84,85.0%,88.8% and 60.5% respectively.The IDI was 0.056(Z test,P<0.05),and there was statistically significant difference in ROC(Delong test,P=0.01).Conclusions The deep learning model trained on the data set of routine CT scans can achieve better diagnostic efficiency in target CT scans data.

关 键 词：体层摄影术 X线计算机 深度学习 肺结节 靶扫描 

分 类 号：R73[医药卫生—肿瘤]